Estimation of domain discontinuities using Hierarchical Bayesian Fay-Herriot models
Section 2. The crime victimization survey
The Dutch crime victimization survey (CVS) is a long-standing survey conducted by Statistics Netherlands at an annual frequency with the purpose to publish reliable figures about crime rates, safety feelings, and satisfaction about police performance in the Netherlands. The CVS is designed to provide reliable figures at the national level and at the level of police districts, which is a subdivision of the Netherlands in 25 regions. The CVS is based on a stratified simple random sampling design for people aged 15 years or older residing in the Netherlands. Strata are formed by police regions to control the precision of these planned domain estimates. The sampling frame is based on the Dutch government’s register of all residents in the Netherlands, called Municipal Basis Administration. The yearly sample of the regular CVS is designed such that about 19,000 respondents are observed. The sample is equally divided over the strata, such that about 760 observations are obtained in each stratum. The general regression (GREG) estimator (Särndal, Swensson and Wretman, 1992) is used to estimate population parameters at the national level and for police districts.
The CVS has been redesigned in 2008. The data collection changed from a mixed-mode design via computer-assisted personal interviewing (CAPI) and computer-assisted telephone interviewing (CATI) to a sequential mixed-mode design that starts with web interviewing (WI) and is followed up for nonrespondents with CAPI and CATI. In addition the questionnaire is changed to improve the wording as well as the order of the questions. To quantify discontinuities induced by this redesign, the regular survey used for official publication purposes was conducted in parallel with the alternative survey approach with a sample size of about 6,000 respondents. In this application, the regular approach was based on the new survey design using WI, CATI and CAPI and the alternative approach was based on the old design using CAPI and CATI data collection only. The sample design for the parallel run is based on stratified simple random sampling where police districts are the strata, using proportional allocation. This results in a sample design that is optimal to estimate figures at the national level but suboptimal for domain estimation.
This survey reports on many different outcome variables. In the present study five key survey variables are considered, see Table 2.1. Estimates for these variables at the national level under the regular and alternative survey are specified in Table 2.2. The sample size allocated to the alternative approach is sufficiently large to estimate discontinuites at the national level using the GREG estimator but insufficient to estimate discontinuities at the domain level of the 25 police districts. The direct estimates for the discontinuities at the national level are indeed significantly different from zero, contrary to the unweighted averages of the direct domain estimates and their standard errors. To obtain more precise predictions for domain discontinuities a model-based small area estimation method based on area level models (Fay and Herriot, 1979) is proposed in the next section.
| Variable | Description |
|---|---|
| Perceived nuisance in the neighborhood on a ten point scale; this includes nuisance by drunk people, neigbours, or groups of youngsters, harassment, and drug related problems. | |
| Percentage of people feeling unsafe at times. | |
| Percentage of people saying to have been victim to property crime in the last 12 months. | |
| Total number of offenses per 100 people. | |
| Percentage of people satisfied with police at their last contact (if contact in last 12 months). |
| Variable | Average over 25 police districts | National level | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| regular | alternative | regular | alternative | |||||||||
| 42.29 | (4.73) | 33.28 | (5.73) | 9.01 | (7.69) | 43.79 | (1.07) | 34.09 | (1.04) | 9.70 | (1.49) | |
| 24.38 | (2.03) | 19.86 | (2.87) | 4.52 | (3.57) | 25.07 | (0.44) | 20.48 | (0.52) | 4.59 | (0.68) | |
| 1.61 | (0.11) | 1.28 | (0.13) | 0.33 | (0.17) | 1.67 | (0.02) | 1.34 | (0.02) | 0.33 | (0.03) | |
| 60.61 | (4.23) | 55.58 | (6.88) | 5.04 | (8.21) | 59.88 | (0.92) | 55.10 | (1.25) | 4.78 | (1.55) | |
| 12.55 | (1.60) | 9.78 | (2.19) | 2.78 | (2.77) | 13.02 | (0.36) | 10.32 | (0.39) | 2.70 | (0.53) | |
- Date modified: